AI for Business Decisions: What to Let It Do, What to Check

A food truck is 72 hours from losing its permit. Picture the owner typing the problem into a chatbot after closing.

The answer arrives fast. It is calm, well formatted and sure of itself. It may also contain a number the owner never gave it.

That food truck was one of ten test situations in a blind experiment. Without rules, seven answers in ten contained an invented fact. With a rulebook attached, one in a hundred and twenty did. Counted fact by fact, it was more than 100 to one.

The machine was the same in both cases. What changed was what it was allowed to do.

The risk isn't the speed of the answer. It's a confident number you never gave the machine. The Jirgenson Model answers with a control system: you set the goals in order, and the AI works from real numbers. It is scientifically proven that the Jirgenson Model works (doi.org/10.5281/zenodo.23964934).

Why a fluent answer is not a safe one

An assistant asked about your business has to say something. So it does.

If a number is missing, it fills the hole with something plausible. If you did not say what matters most, it chooses for you. Then it hands you the result in perfect sentences, and there is nothing to check it against.

That is the trouble with most AI business advice. The grammar is flawless. The source of each number is unknown.

The common enemy here is not the tool. It is the confident guess: a figure that sounds right, arrives fast and was never yours.

AI for business decisions: four jobs, four risks

Any tool, from a spreadsheet to a chatbot, can sit in your control loop in one of four places. Each place has its own way to fail and its own safeguard. Here they are, with Andrew's coffee shop next to each. The last column is an illustration built on his numbers.

Read the table from top to bottom. The further down, the more a mistake costs. So the access you give the assistant should shrink as you go down.

The check-it rule, filled in

You cannot check every answer. You also cannot trust every answer. One line settles it.

Check an answer by hand when p × L is larger than c.

• p is the share of the assistant's answers that turn out wrong on tasks like this one. Take it from your own log.
• L is the cost of the decision if the answer is wrong.
• c is the cost of checking.

Here is the rule at work, with p = 0.1 (one wrong answer in ten):

Then add one test that takes two minutes. Give the assistant your goal rows and your state rows, and ask for the gaps. Andrew's five rows must come back as −1.5, −6, −15, +3 and +25. Any other number means the assistant did not calculate from your data.

Treat that as a red light. Find the cause: the data, the calculation or the version of the rules. Then start a fresh conversation and enter your goals and your state again.

That is the whole tool: four jobs with their safeguards, one line for when to check, and one two-minute test.

What changed when the same AI had rules

Back to the ten test situations. The biggest gains came where an owner is most likely to panic. The translation agency losing clients to automatic translation gained 51 points. The food truck 72 hours from losing its permit gained 42.

Across all ten, the same AI scored 58% higher with the rulebook than with nothing. Here is the comparison that matters to anyone who already writes good prompts:

That is 31% above a good generic prompt. A blind judge preferred the answer with the rules 30 times out of 30. Rule violations fell to zero in 120 blind reviews. Control quality doubled, from 4.2 to 8.3 out of 10. And the answers got shorter, from 501 to 454 words.

One detail is almost funny. The only invented fact in 120 reviews was copied from the prompt's own example sentence. The assistant did not make it up. It borrowed it.

What six months of trusting the guess cost

Andrew's own story shows the cost without any AI in it. He believed his margin was about twenty per cent. It was 9%. That week he had planned a second barista and an advertising campaign, both resting on a margin that did not exist.

Feed that twenty per cent to an assistant, and it will build a handsome plan on it. Fast, tidy and wrong. Within a few months the money runs out, and no one understands why.

With a plan built on the honest 9%, the café reaches two months of reserve at month 10. Left as it was, at month 16. Six months of difference, decided by what you typed in.

"Can I trust a test run by AI?"

Ask that question. The answer is in the limits. Ten scenarios, one model family, an AI judge; the full record of every round, with every answer and judge record, is public on Zenodo. Results with other assistants may differ.

The rule above does not ask you to trust the test. It asks you to keep your own log and to check the expensive answers by hand.

Where the check stops

Be clear about what the check-it rule does. It treats a symptom: bad answers no longer slip through on expensive decisions.

The cause is still in place. The assistant invents and chooses because you have not given it two things. It has no order of goals, so it picks your priorities. It has no thresholds, so it cannot tell a yellow number from a red one.

Two steps close that gap.

• The goal list in refusal order. What you would give up first and what you defend last. Chapter 1.
• The zone thresholds and the red protocol. Where each number turns yellow or red, and what goes first when several are red. Chapter 4.

The free rulebook files, and how to load them step by step, are in the ChatGPT prompt for business. Without your goals and thresholds, even the best ChatGPT prompt for business is guessing about the two things only you know.

Two ways to use ChatGPT for small business

You have two choices. Keep asking the chatbot open questions and act on whatever sounds most confident. Or give it your goals, your numbers and a rule for what to check.

The check-it rule keeps the expensive mistakes out. The refusal order and the thresholds stop the assistant from guessing in the first place.

Test the logic for free: Chapter 1, with the refusal test. Or get the full control system on Kindle, on Amazon. It comes as a book, with the canon and the prompt printed as appendices.

Published research

It is scientifically proven that the Jirgenson Model works. The published experiment, its record and its data:

• Zenodo (DOI): 10.5281/zenodo.23964934
• Hugging Face: E5 Jirgenson Model blind test
• Kaggle: E5 blind test dataset
• GitHub: experiment folder
• Wikidata: dataset Q142022866, author Q141981336
• ORCID: 0009-0001-3034-3552